【摘要】 模型量化技术通过将高精度浮点数映射到低比特离散空间，能有效降低模型存储与计算开销。如何合理化地考虑参数分布的特点以构建更好的映射方案是模型量化研究的核心。现有后训练量化（PTQ）方案多默认非激活层数据呈对称钟型分布，却忽略了激活层与输入引入的微小偏置可能导致的分布不对称性，进而使量化映射偏向一侧，引入显著近似损失。本文针对图像超分辨率（SR）量化方案进行研究，对性能获得公认的二阶段PTQ方案进行改进。将预寻找量化上下界过程中的基于最值的均等分配改进为基于排序的非均匀分配，并在伪量化过程中引入偏置量，将数据的一部分及均值进行自适应偏移，缓解数据偏置引发的估计损失。改进后的方案在效果上全面超越原方案，并同样具有高压缩比以及加速比：相比原SwinIR-light模型参数量减少60%以上，超分辨率过程加速3倍以上。更多还原

【Abstract】 Model quantization technology effectively reduces model storage and computational overhead by mapping high-precision floating-point data to low-bit discrete spaces.A core focus of model quantization research is how to rationally account for the characteristics of parameter distributions to construct superior mapping schemes.Existing Post-Training Quantization(PTQ)schemes nearly universally assume that the data distribution of nonactivation layers follows a symmetric bell-shaped curve,but overlook the fact that small biases introduced by the model′s activation layers and inputs induce distributional asymmetry.Consequently,the resulting quantization mapping is skewed to one side due to this subtle asymmetry,leading to significant approximation loss.This paper investigates quantization schemes for image Super-Resolution(SR)and proposes improvements to the widely recognized two-stage PTQ scheme.First,the max-min-based equal partitioning employed in the pre-search for quantization bounds is modified to a sorting-based non-uniform partitioning approach.Second,a bias term is introduced during the pseudo-quantization process,where a portion of the data and its mean are adaptively shifted to mitigate estimation loss caused by data bias.The improved scheme outperforms the original counterpart across all performance metrics while maintaining comparable high compression ratio and acceleration ratio—compared to the original SwinIR-light model,it reduces the parameter count by more than 60% and accelerates the SR process by more than 3 times.更多还原